Version 1.1: Translation, plus ChatGPT consolidation.
Since 1.0 arrived in June there have been six releases — bug fixes and small feature updates, most of them shaped by people writing in to tell us what was getting in their way. The reception has been warmer than we had any right to expect, and we're genuinely grateful for it. Version 1.1 is the first release with two larger additions: translating foreign-language audio into English, and ChatGPT as a new option for the consolidation step.
Transcribe, translate, and the difference between them
Harvestry has always transcribed — turning speech into text in the same language it was spoken in. A French lecture produced a French transcript. That's still what happens by default, and it's included in both Standard and Pro.
What's new is translate: a French lecture, a Japanese conference talk, a German seminar, rendered directly into an English transcript in a single pass. It isn't a second step bolted onto the end — the model goes from foreign-language audio to English text itself, on your Mac, with no cloud service and no audio leaving the machine. Everything downstream behaves normally: screenshots, timestamps, audio sync, highlights, and consolidation all work on the English result.
Two honest caveats before you buy. The first is that quality is not uniform across the languages Whisper supports. French, Spanish, German, Italian, Portuguese, Dutch and other widely spoken languages translate very well. Low-resource languages — Maltese, Amharic, Shona, Lao, Yoruba and similar — are supported, but the results are markedly weaker, and the failure mode is fluent-sounding English that doesn't match what was actually said rather than obvious nonsense. If your material is in one of those, test it during the trial before committing. Everything runs locally and costs nothing to re-run.
The second is that translation runs into English. That's a constraint of the speech model itself, not a choice we made — it was trained to translate into English and no other target language. French to English works beautifully. French to German is not something the transcription stage can do.
But that pairing isn't out of reach — it just happens one step later. Version 1.1 also adds a Notes language setting for consolidation (Pro). Transcribe a French lecture in French, set Notes language to German, and the consolidation model writes your study notes in German. Because it works from the text rather than the audio, it isn't bound to English, and for less common languages it tends to do better than speech translation does. You'll find it under Settings → Consolidation, and the documentation walks through an example.
Alongside this, you can now tell Harvestry what language is being spoken instead of letting it guess. Detection samples only the opening of the audio, so a video that starts with a music bed, a silent title card, or an English introduction over a French programme can send it down the wrong path — and once it commits, the whole transcript follows. Setting the language explicitly takes one click and removes the guesswork. Both options live in the same menu on the Transcription step, and the export is regenerated if you change your mind after processing.
Consolidation, now with ChatGPT
Consolidation is the optional step that takes your raw, verbatim transcript and asks a language model to turn it into organised notes — headings, structure, the argument pulled out of the digressions. You get both documents in the export: the faithful record and the refined version, side by side.
Until now that meant the Claude API, or a local Ollama model if you have Pro. Version 1.1 adds OpenAI's ChatGPT as a third option. Paste your own API key, pick a model, and it works exactly like the others — the model list is fetched live from OpenAI, so new releases appear in the picker without waiting for a Harvestry update. Switching providers on an already-processed lecture offers to regenerate the notes and re-export, so you can compare how different models handle the same material.
As before, this is the one step that leaves your Mac, it only runs if you turn it on, and it uses your own API key and your own account. If you'd rather nothing left the machine at all, local Ollama consolidation remains the Pro answer.
Who gets what
Transcription is in both versions. Speech to text in the language it was spoken — English to English, French to French, and every other language the model supports — is part of Standard and always has been. Choosing the spoken language explicitly is included too.
Translation to English is a Pro feature. If you're on Standard you'll see the option with a lock beside it, and nothing changes about the transcription you already rely on.
What's in v1.1
- Translate to English (Pro).Turn foreign-language audio directly into an English transcript in one on-device pass — no cloud service, no audio leaving your Mac, and every downstream feature working as usual.
- ChatGPT consolidation.Use OpenAI models for the consolidation step with your own API key, alongside the existing Claude API and local Ollama options. The model list is fetched live, so new releases show up on their own.
- Notes in any language (Pro).Have the consolidation model write your study notes in a different language from the transcript — French to German, Japanese to Spanish, any pair the model knows. Set it under Settings → Consolidation → Notes language.
- Spoken-language selection.Pick the source language yourself instead of relying on auto-detection, which only samples the opening of the audio and can be thrown by intros, music, or silence. Included in both versions.
- Regenerate after changing your mind.Changing the language, the translation setting, the transcription model, or the consolidation provider on a finished lecture now offers to reprocess and re-export, keeping your existing notes and annotations intact.
- Six releases of fixes since 1.0.Faster library loading, more reliable iCloud sync, steadier exports, and a long tail of smaller corrections — nearly all of them reported by people who took the time to write in.